RNA Foundry 为晚期胰腺癌患者 Yash 开源个性化癌症疫苗设计

rnafoundry.com/yash, Made at the RNA Foundry.

In January 2025, 32-year-old Yashdeep Bindal was diagnosed with stage IV pancreatic cancer. He had just returned to work after the birth of his daughter, Maya.

At the RNA Foundry, we are developing a personalized cancer vaccine for Yash. We’re sharing his tumour sequencing data and our vaccine design so that other researchers can help refine it.

Design #2 is our lead vaccine design for Yash. It was generated using matched-normal white-blood-cell DNA alongside tumour DNA and RNA. We share the code, methods, results and sequencing-data access instructions so others can reproduce the work and explore alternative designs.

Design #2 combines 12 mutation, 5 splice, 5 CTA/self and 2 fusion regions. Independent paired Mutect2 discovery, direct normal evidence and tumour-RNA reconstruction support its mutation selection. All 12 eligible mutation regions are included, with no top-N expression-weighted cutoff. Seven original mutation targets were detected in normal DNA and removed; AQP3 and TGFBR2 regions were added. Patient-supported SEC/MITD routing sequences and jointly optimized order/linkers form the protein, which is synonymously encoded with the recorded PopRNA settings. PolyA and any downstream extension are unspecified; nucleotide files contain known upstream segments, not complete molecular sequences. This is one research approach; alternative analyses, antigen selections and designs are welcome.

See the lead design

• Selected antigens and sequences.
• Final protein, antigen order and component coordinates.
• Known RNA segment, coding DNA and annotated DNA segment.
• Methods, every-antigen review and comparison with Design #1.
• Junction check, hydropathy screen, nucleotide verification and completion audit.

FASTA and TSV files are plain text. JSON records inputs, methods and checks. The individual evidence review retains sparse RNA support, RNA-source uncertainties and the encoder's unmet soft folding target alongside the checks that passed.

Verify and reproduce

Use Python 3.11 or 3.12, from the repository root:

python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[mrna]'
yash-vaccine verify
yash-vaccine reproduce --output work/design-2-replay

The first command checks Design #2’s recorded evidence and annotations. The second replays selection, joint order/linker optimization from recorded costs, independent junction checks and synonymous encoding. It compares the selected, protein and nucleotide outputs byte for byte. Neither downloads patient reads, runs new model inference or starts cloud jobs. Repeating raw-read processing or model inference is a separate scope; the Design #2 guide records the Modal controllers, dependencies and large-data locations.

yash-vaccine reproduce defaults to the lead Design #2. Use yash-vaccine reproduce designs/initial for Design #1. Its guide and the usage guide explain the earlier workflow. yash-vaccine compare designs/initial designs/design-2 compares the two complete proteins by stable region identity.

Make another candidate

yash-vaccine init my-candidate

This creates a README and a starter run.py in designs/my-candidate/. Describe your question and implement your approach using any code or tools you need. Reuse our data and helpers where useful, or develop a different workflow. Then run your script as described in its README, for example:

python designs/my-candidate/run.py

Keep your results with your notes: an antigen table, analysis, cassette or whatever you produced. There is no required recipe or pipeline. A design might contain:

designs/my-candidate/
  README.md          Question, methods, inputs, commands and findings
  run.py             Your analysis (or another script/workflow)
  selected.tsv       Chosen antigens, if produced
  cassette.fasta     Construct, if assembled

To change the lead Design #2, use its methods, candidate pool and selected regions as starting points for your own script. reproduce checks the recorded lead; it is not a configurable runner for a changed antigen set.

The optional Design #1 recipe runner varies the earlier tumour-only selection and internal cassette. It does not use Design #2's matched-normal evidence or produce its complete SEC/MITD and RNA stages.

Have a vaccine design for Yash? Submit a pull request. Fork the repository, create a branch and add your work under designs/. Include your selected antigens and rationale, sequences where produced, code and reproduction instructions, and results. Label unspecified components explicitly. The contribution guide explains what to include and how to submit it.

Find the data and methods

• Clinical reports: Stanford pathology/IHC and STAMP, Caris tumour profiling, and Guardant360 liquid biopsy PDFs.
• Data: what is included, what is stored externally, and how to obtain it.
• Matched-normal DNA and Design #2 data: final normal BAM, scoped genotypes, paired calls, QC and exact references; original normal FASTQs excluded.
• Design #2 methods: scientific decisions and evidence behind the lead design.
• Design #1 methods: the earlier design generated without matched-normal DNA.
• Usage: reproduce our candidate or develop your own approach.
• Contribute a design: what to include in a design pull request.
• Upstream stages: DNA/RNA discovery and native scoring contracts.
• AGENTS.md: a short map for coding agents.

Other directions to explore

The shared data support many approaches beyond these designs:

• MHC class II: class-II presentation and CD4 T-cell targets.
• Immunogenicity: predictions of T-cell responses alongside binding scores.
• Processing and presentation: antigen order and linkers optimized for intended target presentation, alongside the current unwanted-junction binding objective.
• Antigen trafficking: compare the current SEC–antigens–MITD architecture with other routing elements or a cassette without them.

Repository map

Provenance maps the input, model and verification records. Large sequencing files and case-analysis outputs are available through a public read-only endpoint. Download instructions and the file catalog describe the files, sizes and checksums. Third-party tools, models and references come from their providers. Full upstream RNA modules are not bundled; local replay uses included evidence and recorded predictions. Required credits and retained-source notices cover the materials used here.